ml4ai/mliis

Code for meta-learning initializations for image segmentation

Meta-Learning Initializations for Image Segmentation

Code for meta-learning and evaluating initializations for image segmentation as described in our paper https://arxiv.org/abs/1912.06290, which was presented at the 4th Workshop on Meta-Learning at NeurIPS 2020.

Note that this repository is in archival status. Code is provided as-is and no updates are expected.

Example 5 shot predictions on test samples from meta-test tasks:

Citing

If you find this project useful in your research, please consider citing:

@article{hendryx2019meta,
  title={Meta-Learning Initializations for Image Segmentation},
  author={Hendryx, Sean M and Leach, Andrew B and Hein, Paul D and Morrison, Clayton T},
  journal={4th Workshop on Meta-Learning at NeurIPS 2020},
  year={2020},
}

Setup

We have included a requirements.txt file with dependencies. You can also see make_python_virtualenv.sh for recommended steps for setting up your environment.

You can download the FSS-1000 meta-training and evaluation tfrecord shards from: https://drive.google.com/open?id=1aGHP0ev_1eAFSnYtN0ObDI-DnB0TsQUU

And the joint-training shards from: https://drive.google.com/open?id=1aQpyQ0CEBCL9EW8xoCaI6xveYxtXNYKq

The FP-k dataset shards are available at: https://drive.google.com/open?id=1G1NJIyQlkxAb4vlsRDPR3W3If_RJ4rPd

The FP-k dataset is derived from the FSS-1000 and PASCAL-5i datasets. PASCAL-5i was in turn derived from the parent datasets: PASCAL and Semantic Boundaries Datasets as described in One-Shot Learning for Semantic Segmentation .

We created our meta-training tfrecord shards by following these steps. Download the FSS-1000 dataset from https://github.com/HKUSTCV/FSS-1000 Convert the images and masks to tfrecords:

python fss_1000_image_to_tfrecord.py --input_dir <path to images and masks> --tfrecord_dir <directory to write tfrecords in>

Run the SOTA evaluation

Extract the checkpoint:

tar -xzvf EfficientLab-6-3_FOMAML-star_checkpoint.tar.gz

Put the FSS-1000 meta-training and evaluation tfrecord shards at the root of this repo or edit the data_dir path in run.sh to point to the shards on your machine.

Finally, call:

./run.sh

Run an experiment

The main point of entry in this codebase is:

python run_metasegnet.py <args>

See args.py for arguments and their descriptions.

Our SOTA meta-learned initialization that generated the best FSS-1000 results reported in our paper is in this repository at EfficientLab-6-3_FOMAML-star_checkpoint

Visualize predictions

To see predictions, set the environment variable ala:

export SAVE_PREDICTIONS=1

Save an adapted model

To save the weights of an updated model on your task(s), run:

python run_metasegnet.py --save_fine_tuned_checkpoints --save_fine_tuned_checkpoints_dir /path/to/save/to <--other_args>

See run.sh for our recommended hyperparameters found via update hyperparameter optimization.

EfficientLab

Our SOTA network architecture class is defined in models/efficientlab.py.

Acknowledgements

This repository builds on the Reptile implementation by OpenAI and the EfficientNet backbone implementation by Google.

mliis's Issues

how can I use muti-gpu training? I only can use CPU.

Thanks very much!

How much time did you spend training this network?

This work is awesome!
I'm training this model on FSS-1000, but I find the training is slow and can not use multi gpus.
So I want to konw how much time did you spend training this network and what gpu did you use?
It‘s helpful to determine if I reproduced the model incorrectly.
Thanks very much!